About the author
Yesterday, 6 November 2018, the AMD Next Horizon conference was held in San Francisco7, dedicated to the upcoming server solutions of this brand. And no, the "seven" in the name of the event does not specify any dating, nor is it more coincidental. The whole event took place around the 7 nm lithographic process and the products produced there, namely the EPYC Rome (Zen 2) and the Radeon Instinct accelerators with the Vega 20 chip.
In addition to product announcements, a real investor bomb was also prepared: AMD and Amazon announced a partnership under which
EPYC processors will feed the cloud infrastructure of Amazon Web Services. The exchange reaction did not have to wait long.
Zen 2 or eight cores in the cluster
The transition to 7 nm technology will begin with different styles than AMD will normally work with Zen processors. While so far the publishing cycle has begun with Ryzens consumer, now the first EPYC Rome systems will appear. TSMC, not GlobalFoundries, will have a production weight to date, and the new slogan is said to be double: half the power consumption and one-fourth best performance per watt.
Obviously, such increased compression must bring visible changes to the processor's front-end. Most importantly, the CCX clusters, the smallest component of Zenów, will reach eight cores. Thus, AMD can double the number of cores in each segment without increasing the physical size of the chips. The new EPYC Rome will offer them 64, while its predecessor from the Naples family has 32. With the same principle, consumer Ryzen could now have 16 cores, though that is only a guess.
Nevertheless, with the application of the new microarchitecture, the manufacturer promises much more changes:
- improved execution pipe,
- 256-bit mobile unit,
- changes in branch prediction,
- better support for original downloads,
- increased command buffer,
- optimization of the sender's mechanism.
Eventually, EPYC Rome will twice guarantee the performance of its predecessor by looking around the clock and up to four times to improve efficiency in mobile computing.
And as I have already mentioned, this system will be produced in 7 nm technology, if only in part. The CCX clusters, where the kernels are located, will be created in the last generation of the lithographic process. However, chip logic, including elements such as eight DDR4 controllers, PCI Express 4.0 interfaces and interfaces, will be created in current 14 nm technology. Either way, AMD can beat the win. In the server segment, Intel will drop 14 nm probably in 2020 alone, along with Ice Lake-SP-10 nm units. Even the recent Cascade Lake-AP communiqué, dating back to 2019, offers nothing new in lithography.
The first graphics processor at 7 nm
As expected, the first AMD graphics processor in 7 nm technology is Vega 20, so shrink which is nominally produced in 14 nm Vega 10. The GPU has 13.23 billion transistors at 331 mm² instead of 12.5 billion and 510 mm2 respectively. It is also known to have slightly higher clocks, which result from the performance in FP32 precision floating point calculations, at 14.8 TF, with 12.6 TF.
The senior vice president of engineering, David Wang, announced that Vega 20 will not appear as a product for players. Instead, they will activate the Radeon Instinct M60 learning accelerator, available as a PCI-E 4.0 expansion card, with 32 GB of HBM2 memory (bandwidth: 1 TB / sec).
Radeon Instinct M60, as opposed to Vega 10 based products, will also receive an INT8 and INT4 command set, as well as MxGPU support for virtualization. I am talking here about a product exclusively dedicated to supercomputers (HPC – High Performance Computers). yet shrink Vega, from the point of view of the home user, should determine the path of the next GPU at 7 nm, with time-consumer.
Especially that the slightly cheaper model M50, also equipped with 3840 shading units instead of 4096 and 16 GB of HBM2, is also planned. Then he could hit the players.